How to Use KEDA and Prometheus to Scale your Kubernetes Workloads

These days, everyone and their grandma are using Kubernetes and one important aspect of Kubernetes is scaling your workloads. With KEDA, it is extremely simple to scale your workloads! Let’s have a look.

repository: https://github.com/djamaile/keda-demo

Introduction to KEDA

Straight from the website of KEDA:

KEDA is a Kubernetes-based Event Driven Autoscaler. With KEDA, you can drive the scaling of any container in Kubernetes based on the number of events needing to be processed.

KEDA provides many 'triggers' on which your application can scale on. For example, Prometheus, PubSub, Postgres and many more. In this blog post we will focus on Prometheus.

Starting up

First, let's spin up a cluster! I am using kind but you are free to use minikube if you prefer that :).

$ kind create cluster

Create the namespace

$ kubectl create ns keda-demo

Switch to the namespace

$ kubectl config set-context --current --namespace=keda-demo

If the cluster is spun up, we can start deploying our Prometheus. For this, I have already written a prometheus manifest so you won’t have to do it.

prometheus.yaml

apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: prometheus
rules:
  - apiGroups: [""]
    resources:
      - services
    verbs: ["get", "list", "watch"]
  - nonResourceURLs: ["/metrics"]
    verbs: ["get"]
---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: keda-demo
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: prometheus
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: prometheus
subjects:
  - kind: ServiceAccount
    name: keda-demo
    namespace: keda-demo
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: prom-conf
  labels:
    name: prom-conf
data:
  prometheus.yml: |-
    global:
      scrape_interval: 5s
      evaluation_interval: 5s
    scrape_configs:
      - job_name: 'go-prom-job'
        kubernetes_sd_configs:
        - role: service
        relabel_configs:
        - source_labels: [__meta_kubernetes_service_label_run]
          regex: go-prom-app-service
          action: keep
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: prometheus-deployment
spec:
  replicas: 1
  selector:
    matchLabels:
      app: prometheus-server
  template:
    metadata:
      labels:
        app: prometheus-server
    spec:
      serviceAccountName: keda-demo
      containers:
        - name: prometheus
          image: prom/prometheus
          args:
            - "--config.file=/etc/prometheus/prometheus.yml"
            - "--storage.tsdb.path=/prometheus/"
          ports:
            - containerPort: 9090
          volumeMounts:
            - name: prometheus-config-volume
              mountPath: /etc/prometheus/
            - name: prometheus-storage-volume
              mountPath: /prometheus/
      volumes:
        - name: prometheus-config-volume
          configMap:
            defaultMode: 420
            name: prom-conf

        - name: prometheus-storage-volume
          emptyDir: {}
---
apiVersion: v1
kind: Service
metadata:
  name: prometheus-service
spec:
  ports:
    - port: 9090
      protocol: TCP
  selector:
    app: prometheus-server

The Prometheus manifest is really simple. Just a Prometheus workload with a clusterrole and a clusterrolebinding. Don't forget to apply the manifest:

$ kubectl apply -f prometheus.yaml

Once the pod is up and running, let's see if it also works:

$ kubectl port-forward svc/prometheus-service 9090

Now visit http://localhost:9090 and you should see the user interface of Prometheus.

Deploying Keda

We can now deploy the KEDA operator. KEDA provides multiple ways to deploy their operator, but for now we will use the k8s manifest.

$ kubectl apply -f https://github.com/kedacore/keda/releases/download/v2.4.0/keda-2.4.0.yaml

Now there should be two pods in the namespace keda you can check it with the following command:

$ kubectl get pods -n keda

As you can see there are two pods being spinned up:

on 🤠 kind-kind (keda) Desktop/projects/keda-prometheus ☁️  default
🕙[ 07:35:40 ] ❯ kubectl get pods                                                         335ms
NAME                                      READY   STATUS              RESTARTS   AGE
keda-metrics-apiserver-66b8c68649-2mwf8   0/1     ContainerCreating   0          5s
keda-operator-574c6d4769-q9mlc            0/1     ContainerCreating   0          5s

The metrics-apiserver exposes data to the Horizontal Pod Autoscaler, which gets consumed by a deployment. The operator pod activates Kubernetes deployments to scale to and from zero on no events.

Creating the application (Optional)

The application is a simple go application that increments the metric http_requests when you visit it. This section is optional because you are also free to use my docker image.

in your folder execute the following:

go mod init github.com/djamaile/keda-demo

Then in your main.go you can put in the following code:

package main

import (
	"fmt"
	"log"
	"net/http"

	"github.com/prometheus/client_golang/prometheus"
	"github.com/prometheus/client_golang/prometheus/promhttp"
)

type Labels map[string]string

var (
	httpRequestsCounter = prometheus.NewCounter(prometheus.CounterOpts{
		Name: "http_requests",
		Help: "number of http requests",
	})
)

func init() {
	// Metrics have to be registered to be exposed:
	prometheus.MustRegister(httpRequestsCounter)
}

func main() {
	http.Handle("/metrics", promhttp.Handler())
	http.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
		defer httpRequestsCounter.Inc()
		fmt.Fprintf(w, "Hello, you've requested: %s\n", r.URL.Path)
	})
	log.Fatal(http.ListenAndServe(":8080", nil))
}

Now build the go application with:

$ go mod tidy

Let's then make a simple Dockerfile for it:

FROM golang as build-stage

COPY go.mod /
COPY go.sum /
COPY main.go /
RUN cd / && CGO_ENABLED=0 GOOS=linux go build -a -installsuffix cgo -o go-prom-app

FROM alpine
COPY --from=build-stage /go-prom-app /
EXPOSE 8080
CMD ["/go-prom-app"]

Only thing left is to build and push the image:

$ docker build -t <your_username>/keda .
$ docker push <your_username>/keda

Running the application

If you don’t have a Docker account or don’t want to use it, that’s fine. You can use my docker image! Let’s get our go application running in the cluster, for that we need some k8s manifests. Not to worry because I already wrote them:

go-deployment.yaml

apiVersion: apps/v1
kind: Deployment
metadata:
  name: go-prom-app
  namespace: keda-demo
spec:
  selector:
    matchLabels:
      app: go-prom-app
  template:
    metadata:
      labels:
        app: go-prom-app
    spec:
      containers:
        - name: go-prom-app
          image: djam97/keda
          imagePullPolicy: Always
          ports:
            - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: go-prom-app-service
  namespace: keda-demo
  labels:
    run: go-prom-app-service
spec:
  ports:
    - port: 8080
      protocol: TCP
  selector:
    app: go-prom-app

You can replace the image name with your own image if you prefer that. Let's apply the manifest:

$ kubectl apply -f go-deployment.yaml

If the pod is up verify if it is working

$ kubectl port-forward svc/go-prom-app-service 8080

If you visit http://localhost:8080 you should see Hello, you've requested: /.

Scaling the application

Now that we have our go application up we can write a manifest that will scale our application. Keda offers many triggers that can scale our application, but of course, we will use the Prometheus trigger.

In a new file called scaled-object.yaml add the following content:

apiVersion: keda.sh/v1alpha1
# Custom CRD provisioned by the Keda operator
kind: ScaledObject
metadata:
  name: prometheus-scaledobject
spec:
  scaleTargetRef:
    # target our deployment
    name: go-prom-app
  # Interval to when to query Prometheus
  pollingInterval: 15
  # The period to wait after the last trigger reported active
  # before scaling the deployment back to 1
  cooldownPeriod: 30
  # min replicas keda will scale to
  # if you have an app that has an dependency on pubsub
  # this would be a good use case to set it to zero
  # why keep your app running if your topic has no messages?
  minReplicaCount: 1
  # max replicas keda will scale to
  maxReplicaCount: 20
  advanced:
    # HPA config
    # Read about it here: https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/
    horizontalPodAutoscalerConfig:
      behavior:
        scaleDown:
          stabilizationWindowSeconds: 30
          policies:
            - type: Percent
              value: 50
              periodSeconds: 30
        scaleUp:
          stabilizationWindowSeconds: 0
          policies:
            - type: Percent
              value: 50
              periodSeconds: 10
  triggers:
    - type: prometheus
      metadata:
        # address where keda can reach our prometheus on
        serverAddress: http://prometheus-service.keda-demo.svc.cluster.local:9090
        # metric on what we want to scale
        metricName: http_requests_total
        # if treshold is reached then Keda will scale our deployment
        threshold: "100"
        query: sum(rate(http_requests[1m]))

Read the yaml manifest and its comments to understand what is going on. One important note as well is in advanced.horizontalPodAutoscalerConfig.scaleUp.policies you can see I have specified 50%, that means our pod will scale up with 50% of it’s current amount of pods. 1 -> 2 -> 3 -> 5 -> 8 -> 12 -> 18 -> 20 it will stop at 20 pods because that is the limit we specified.

Let's apply the manifest:

$ kubectl apply -f scaled-object.yaml

This will provision an HPA in your namespace which you can check with:

$ kubectl get hpa

but because this is a custom CRD you can also query the custom CRD with kubectl:

$ kubectl get scaledobject.keda.sh/prometheus-scaledobject

You should see that the prometheus-scaledobject is ready so let’s scale our application! Remember our application scales on the metric http_requests_total and our threshold is only 100 so we should be able reach that threshold. For this we can use a simple tool called hey.

Run the application

$ kubectl port-forward svc/go-prom-app-service 8080

In another terminal watch the pods

$ kubectl get pods -w -n keda-demo

Put load on the application (Do this continuously, until there are 20 pods)

$ hey -n 10000 -m GET http://localhost:8080

It can take a minute before the application actually starts scaling. After a while you should have 20 pods up and running! Now let’s also look at the scale down process. Stop putting load on the application and let’s just watch the pods. This process should go from 20 -> 10 -> 5 - > 2 -> 1. This is basically how KEDA goes to work!

If you like KEDA please check out their docs for more examples and what type of different triggers they provide. Happy auto-scaling!